New independent preprint in Applied Physics & Signal Processing — `ScarCodec` tested without modifications on 5 official public datasets (NASA + FITS + PhysioNet; zero synthetic data):
• `NASA Spitzer IRAC` FITS: `11.77x` compression (vs `1.14x` for `zlib`), `O(1)` read access that is `32.1x` faster, and a `2,093.6x` reduction in peak stellar error compared to byte-aligned FP16 `rFFT`.
• `NASA Space Shuttle` Telemetry: `6/6 (100%)` of rare valve failures (`0.04%` of data) preserved via `O(1)` access (`Δ = 0 B`), compared to `0/6` preserved using a circular buffer.
• `PhysioNet 64-channel EEG`: `10.47x` compression in a blind cross-run test (`+14.63 dB` SNR; `cos = 0.9967`), reconstructing 32 hidden electrodes within a fixed `34.8 KB` footprint.
Code & SHA-256 replication: https://github.com/byte-od/mixture-of-scars
thetod•8m ago